UrbanShift is a global program that supports cities to adopt integrated approaches to urban development, shaping low-carbon, climate-resilient communities where people and planet both can thrive. The program is funded by the Global Environment Facility (GEF) and jointly managed by a global team consisting of the United Nations Environment Program (UNEP), World Resources Institute (WRI), C40 Cities and ICLEI Local Governments for Sustainability. The initiative supports 23 cities across nine countries, providing the knowledge, tools and training they need to transform their urban fabric and shift towards a more sustainable, equitable future.
As one of the key activities to support development of a knowledge-base for the UrbanShift initiative and all participant cities, the WRI data team will work with UrbanShift cities to identify and provide all cities with a common set of critical spatial data layers. using open source, global data. World Resources Institute is providing several types of data-related assistance to participating cities:
Outputs will include datasets, indicators and replicable analysis methods relevant to all cities. Additionally, analyses customized to the specific themes of interest for each city will be provided. Finally, an UrbanShift Lab will be delivered for which these data and analyses may act as one input.
Freetown City Council (FCC) identified several potential areas of WRI assistance based on ongoing activities and needs associated with the Freetown Structure Plan and the Mayor’s Transform Freetown plan. Assistance on the **Analysis of urban expansion and infill trends* has been identified as a key thematic of analysis in order provide a replicable method using global open data for assessing urbanization trends in the selected cities.
WRI proposes to provide FCC with a supervised classification framework to develop simple models for classifying urbanized land in multiple time periods. The framework provide a starting classification model that can be customized with additional user inputssuch as additional local datasets and land classes labeling based on local knowledge. In addition to urban vs non-urban classification, WRI may be able to provide a model framework for classifying multiple types of urban expansion (e.g., infill, sprawl, leapfrogging, etc.).
In order to build a land classification model in the Freetown region, thre main datasets are needed: administrative boundaries of Freetown, high resolution earth observation imagery (Sentinel-2 product), and labeled data providing samples of urban and non urban boundary areas. The table below gives detailed information about these datasets:
| Dataset | source | Description | Temporal extent | Doc | Download | Format |
|---|---|---|---|---|---|---|
| Administative boundaries | geoBoundaries database | Open data of boundaries for every country in the world at multiple administrative levels provided by geoLab. | From 2017 to present (yearly updates) | Link to doc | Link to download Freetown boundaries | Geojson |
| Sentinel-2 satellite imagery | European Union/ESA/Copernicus | Sentinel-2 is a wide-swath, high-resolution, multi-spectral imaging mission supporting Copernicus Land Monitoring studies, including the monitoring of vegetation, soil and water cover, as well as observation of inland waterways and coastal areas. The Sentinel-2 data contain 13 UINT16 spectral bands: four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution. | From 2015-06-23 to present | Link to doc | Link to download Sentinel-2 data | Geotiff |
| Labeled training data | Manual | Manual definition of Urban and Non urban polygons within Freetown region. | Non urban resgions, Urban regions | Geojson |
The first step consists of collecting raw input data that will be used for building the land classification machine learning model.
The administrative boundaries data are obtained from the geoBoundaries database. Two administrative levels have been used for this analysis:
Sentinel-2 satellite imagery data are obtained from Google Earth Engine Data Catalog. The images can be extracted by filtering based on the area of interest (Freetown region) and time period of analysis. Since Sentinel-2 data are provided from 2015-06-23 to present, six years data have been extracted corresponding to the years: 2016, 2017, 2018, 2019, 2020, 2021. Hence, we can train the model on 2016 Sentinel-2 data and apply it to the other selected years in order to assess urban expansion during the last six years.
A supervised classification is based on the idea that a user can select sample pixels in an image that are representative of specific land classes and use these training sites as references for the classification of all other pixels in the image. These training sets are often selected based on the knowledge of the user or extracted from ground truth data. For this analysis, we selected 53 regions (polygons) within the Freetown area and classified them into urban and non urban classes. These polygons are used as training set and the objective of the machine learning model consists of learning a function that maps the pixels’ characteristics of these regions with the two classes. While this method enables us to build a first version of the land classification model, increasing manually labeled data and ensuring an accurate definition of urban land classes is needed to improve the model robustness.
The Copernicus Sentinel-2 TOA imagery is organised as an ImageCollection object, which is a container for a collection of individual images. The collected images are filtered in order to restrict selected image tiles to Freetown administrative boundaries and within the specified date range. In addition, clouds are masked from each image using their corresponding cloud probability layer. Two functions are used to achieve cloud masking: a function to join the cloud probability layer to the relevant image and one to apply the mask where cloud probability is greater than 50 percent. Finally, a medoid composite is generated from the set of overlapping pixels by selecting the pixel nearest to the multi-dimensional median of overlapping pixels (Flood, 2013). The result minimizes image contamination from residual clouds and cloud shadows. (source)
The labeled data is randomly split (70% / 30%) into training and validation datasets. The training set is used for training the machine learning model and the validation set is used for evaluating the trained model and comparing the model results with ground truth.
We selected a Random Forest (RF) classifier for implementing binary land classification. This classifier is suitable for this kind of analysis and default parameters have been used for proposed model. The output images include values 0 for non-urban areas and 1 for urban areas.
Once the model is trained, we use the validation sample in order to evaluate the reliability of the classification outputs. Accuracy is one metric for evaluating classification models. It corresponds to the percentage of right predictions produced by the model. Formally, it can be considered as the percentage of correct predictions:
\[Accuracy = \frac{NbCorrectPrediction}{TotalPrediction}\] Where:
NbCorrectPrediction = TP (True Positive)+ TN (True Negative)
NbCorrectPrediction = TP (True Positive)+ TN (True Negative) + FP (False Positive) + FN (False Negative)
The obtained trained model can be applied on different years of Sentinel-2 imagery data in order to track the urban area expansion during a time period of interest.
The map below shows the results of the classification model for the years 2016 and 2021. The results highlighted an increase of urban area from 22.18% in 2016 to 29.86% in 2021. The most important part of this urban expansion is situated at the limits of the Western Area Forest Reserve.
In order to localize the increase of urban area, we calculated the difference in urban area pixels between 2016 and 2021. The red pixels highlight urban expansion hotspots:
We focus in this section on the urban expansion at the Freetown municipality level. The map below provides the spatial distribution of the urban and non-urban areas within the Freetown municipality for the years 2016 and 2021.
The outputs of the land classification model highlights an increasing of urban areas of 5.51 %.
| name | year | Urban | Non Urban |
|---|---|---|---|
| Freetown Core | 2016 | 75.61 | 24.39 |
| Freetown Core | 2021 | 81.12 | 18.88 |
This section provides urban expansion indicators aggregated at the ward level. The Freetown municipality contains 48 wards. The map below presents the urban expansion percent by comparing the urban area in 2016 an 2021. The wards presenting the highest urban expansion are Thnderhill and Mamba Ridge II with more than 20% increase of urban areas.
| Ward number | Urban expansion 2016-2021 (%) |
|---|---|
| 411 | 24.78 |
| 418 | 21.60 |
| 410 | 18.11 |
| 419 | 13.01 |
| 406 | 11.70 |
| 434 | 8.60 |
| 403 | 8.39 |
| 400 | 7.90 |
| 412 | 7.61 |
| 426 | 7.31 |
| 446 | 6.46 |
| 428 | 6.37 |
| 399 | 6.15 |
| 414 | 5.73 |
| 442 | 5.45 |
| 441 | 5.36 |
| 440 | 4.97 |
| 415 | 4.84 |
| 437 | 3.89 |
| 404 | 3.82 |
| 405 | 3.46 |
| 425 | 3.25 |
| 402 | 2.91 |
| 416 | 2.82 |
| 435 | 2.48 |
| 444 | 2.37 |
| 438 | 2.33 |
| 429 | 2.27 |
| 407 | 1.90 |
| 433 | 1.77 |
| 417 | 1.69 |
| 401 | 1.69 |
| 423 | 1.63 |
| 436 | 1.63 |
| 443 | 1.61 |
| 420 | 1.52 |
| 413 | 0.86 |
| 430 | 0.44 |
| 408 | 0.42 |
| 439 | 0.28 |
| 421 | 0.15 |
| 431 | 0.14 |
| 409 | -0.11 |
| 432 | -0.12 |
| 445 | -0.48 |
| 427 | -1.13 |
| 424 | -1.32 |
| 422 | -1.74 |
This section provides urban expansion indicators aggregated at the planning area level. The Freetown municipality is divided into 12 planning areas. The map below presents the urban expansion percent by comparing the urban area in 2016 an 2021. The planning areas presenting the highest urban expansion are 8 and 9 with respectively 19.3% and 12.9% increase in urban areas.
| Planning area number | Urban expansion 2016-2021 (%) |
|---|---|
| 8 | 19.33 |
| 9 | 12.91 |
| 2 | 6.57 |
| 11 | 5.66 |
| 5 | 4.88 |
| 10 | 4.49 |
| 7 | 4.15 |
| 1 | 3.42 |
| 3 | 2.29 |
| 6 | 2.10 |
| 4 | 0.33 |
By combining land use data with urban expansion indicator, we can identify the land use types that present highest increase in terms of urbanization. The land use data collected by the FCC classify Freetown municipality area into 13 distinct classes:Civic and Culture, Coastal Wetland, Mangrove, Commercial, Industry, Mix Commercial with Residential, Open Spaces - Recreational - Sport, Residential low density, Residential medium density, Residential high density, Security - Utility, Urban Agriculture, Waterbody - River - Creek, and Woodland and Forest.
The aggregation of urban expansion indicator by land use class shows that the areas with highest urban sprawl are the waterbody and river, woodland and forest, and urban agriculture.
| Land use class | Average Urban expansion 2016-2021 (%) |
|---|---|
| Waterbody, River, Creek - WA | 28.09 |
| Woodland, Forest - FO | 18.39 |
| Urban Agriculture - UA | 14.30 |
| Coastal Wetland, Mangrove - CW | 4.35 |
| Residential, low density - RE-L | 3.92 |
| Residential, medium density - RE-M | 2.63 |
| Civic and Culture - CI | 1.84 |
| Open Spaces, Recreational, Sport - OS | 1.68 |
| Residential, high density - RE-H | 1.35 |
| Industry - IN | 1.10 |
| Commercial - CO | 0.68 |
| Mix Commercial with Residential - CO-R | -0.34 |
| Security - SE, Utility - UT | -1.36 |
These analyses provide a proof of concept of using global remote sensing open data to provide insights on urban expansion in the Freetown region. We proposed a classification machine learning model trained on labeled data for classifying land into urban and non urban areas. We used Random Forest classifier built based on 2016 Sentinel-2 imagery and we applied it on 2021 data in order to assess urban expansion.
The trained model presents a high accuracy (95%) based on the validation set extracted from the labeled data. The application of this model enabled us to explore the spatial distribution of urban and non urban areas in the whole Freetown region at different periods. The outputs show an increase in urban expansion encroaching on the Western Area Forest Reserve.
We proposed an urban expansion indicator based on the percentage of urban area between between 2016 and 2021. Aggregated at the ward level, the proposed indicator highlighted the cities that present the highest urban expansion, such as Thnderhill and Mamba Ridge II (more than 20% increase of urban areas in the last 6 years).
The results of this analysis may be used to better plan prevent urban expansion in the identified areas and localities as part of the development of Freetown master plan.
While the results of the proposed classification model provide insights on urban expansion within Freetown region, we have identified potential ways to improve the model: :
Data labeling methodology: As discussed in the methodology section, the classification model is based on training labeled data to map the dependencies between remote sensing images and the defined classes. For this prototype, only 53 polygons have been used as training set (26 urban vs 27 non-urban). These training data have been defined based on visual/manual identification of urban classes in Freetown. More accurate ground truth data could be collected (from potential existing data sources or local territory knowledge) to improve the model accuracy when applied in new satellite imagery.
More land classes: We use a binary classification model for determining urban vs non-urban area. However, it is possible to include other land classes depending on the availability of labeled data for those additional classes.